Drift

A substance-level difference between a delivered teaching and its source material, such as a dropped qualifier, a misstated timing, or a skipped step, distinguished from acceptable delivery variation, which changes wording or examples without changing what a listener actually learns to do.

The gradual degradation of an AI model as the real world moves away from the data the model was trained on, so a model that was accurate at launch becomes quietly wrong over time. It is the technical reason an AI capability must be tended continuously rather than shipped and abandoned.

A bias that changes over time, driven by temperature variation, supply voltage changes, or component aging. Drift is distinguished from static bias by its dependence on an external variable. Gyroscope temperature drift is a primary limitation of pure inertial dead reckoning.

The gradual degradation of an AI capability as the real-world conditions it operates in diverge from the conditions it was built or trained on. Drift makes the Maintain bucket non-optional; a maintenance estimate near zero assumes the world stops changing. (see Topic 4.6)

Gradual, unintended change in what data means, what a model's inputs look like, or how well a model performs, accumulating without a single triggering event and typically discovered only in hindsight.